How an ITP Maker Used AI to Build a Digital Portrait

When a creative technologist sits down with a blank canvas, the question is rarely what tool to pick but how to make the tool disappear. That tension sits at the heart of a project featured on the ITP 30 Show, where one participant reached for generative algorithms to build a digital likeness that feels strangely alive. The result is part experiment, part portrait, and part quiet argument about authorship in the age of synthesised imagery.

For Australian viewers, the work lands in a moment when local galleries, from the National Portrait Gallery in Canberra to ACMI in Melbourne, are actively debating what counts as a portrait when the sitter is a prompt and the painter is a model. The piece resonates with conversations happening at Vivid Sydney, where light, data and identity are projected onto the harbour walls every winter.

The spark behind the pixels

The project began with a notebook sketch and an itch to see what a machine could remember. The maker wanted to explore whether a face conjured entirely by code could carry the same weight as a charcoal drawing of a friend in Brunswick or a watercolour of a relative in Fremantle. The answer turned out to hinge on what data the system had already absorbed.

Rather than training a fresh model from scratch, the participant leaned on existing systems and treated them as collaborators. That framing shaped lighting decisions and the choice of a slightly off-centre gaze, a compositional trick borrowed from European studio photography. You can read more in the journal entry tracing the evolution.

Choosing the generative engine

Diffusion models, GANs and transformer-based vision systems each offered a different flavour of output. The maker spent two weekends comparing how a face dissolved and reformed across platforms, looking for one that produced skin texture closer to a Mornington Peninsula fog than a CGI sheen. Stability, control and the ability to seed specific features all mattered, since the final image had to be reproducible for a live audience.

Local infrastructure shaped the choice too. Running heavy inference on a laptop in a share house in Fitzroy quickly throttled, so the participant offloaded heavier passes to a cloud runtime, then downloaded results to retouch locally. The hybrid approach mirrors a trend across Australian studios, where artists increasingly route compute through both home setups and rented GPUs to balance cost with creative control.

Prompting as portraiture

Writing a prompt for a synthetic face turned out to be a curious kind of writing. Each adjective altered the mood, each lighting cue nudged the likeness, and every cultural reference carried baggage the model had absorbed from its training set. The participant treated the prompt less as a search query and more as a brief written for a stubborn collaborator who had never set foot in Adelaide or Parramatta.

One passage simply asked for a figure under harsh overhead light, mid-blink, with freckles and a faint scar above the brow. Another asked the same model to render the subject in soft, diffused window light from a Fitzroy terrace. Comparing the two outputs side by side became its own small body of work, a study in how much atmosphere a single sentence can carry.

Iteration, rejection and re-seeding

The first fifty generations were discarded. Faces that looked too symmetrical, expressions that drifted into uncanny territory, and lighting that flattened the subject all went into a rejection folder that soon resembled a casting call in a Surry Hills studio. The participant kept iterating, swapping seeds, tweaking noise schedules, and rerunning generations with subtle adjustments until the geometry stopped feeling algorithmic.

Outpainting tools and inpainting brushes turned the workflow into something closer to traditional editing than pure generation. The maker manually painted over artefacts, restored an ear that had dissolved into hair, and sharpened the edge of a collar against a wall painted in the warm grey common to renovated Sydney terraces. Each manual fix brought the image closer to feeling like a portrait rather than a screenshot.

Putting the portrait on display

For the live showing, the final still image was paired with a slow generative loop that breathed new life into the face every few minutes. Visitors at the showcase could lean toward a small screen and watch hair drift, freckles shift, and ambient light rebalance across the subject's face. The piece sat beside a printed catalogue and an audio interview, all bound into a single illustrated volume.

The format echoed zine culture in Melbourne's inner north, where risograph prints pair digital tools with hand-finished pages. A walkthrough of similar projects sits in the project archive, which collects dozens of experimental works from past ITP cohorts.

How audiences read the face

Reactions split along a familiar fault line. Some viewers immediately asked who the person was, expecting a sitter to step forward from the crowd. Others accepted the image as a portrait of a kind of person, a composite drawn from millions of unnamed faces, and asked instead where the model had been trained and what biases it might carry. Both questions felt welcome to the maker.

The conversation carried parallels with the rollout of AI imaging tools in Australian creative industries, where unions like the Media, Entertainment and Arts Alliance have pushed for disclosure of synthetic assets in advertising. Watching audiences negotiate their trust in real time gave the project a usefulness no printed caption could match.

What it means for Australian practice

Australian studios and universities are absorbing generative imaging quickly, in cities well beyond Sydney and Melbourne. Brisbane's QUT, Perth's Curtin, and RMIT in Carlton have all introduced modules on machine-made imagery within the past two years. The ITP project's open approach to process, showing prompts, seeds and rejected takes, sits comfortably with that local ethic of transparent craft.

It also lands during a surge of interest in synthetic identity across the local market, from fashion campaigns shot with diffusion tools to commercials for Australian banks using AI-assisted casting. Treating a generated face as just another portrait subject may be the most useful move a regional practitioner can make right now. For a taste of similar experiments, the weekly tournament write-up offers a playful counterweight.

If you are working with synthetic tools in your own practice, consider submitting to the next ITP 30 Show, attending open studio events in Sydney and Brisbane, or exploring the archive to see how peers handle authorship, data and display. The work gets sharper when more hands help shape it.